OpenAI's AI Model Discovers Critical PHP Vulnerabilities Affecting Majority of Websites
Researchers using OpenAI's GPT-5.6 Sol uncovered two critical security flaws in PHP that could compromise roughly 70% of internet websites, prompting a coordinated global response to patch the vulnerabilities.
On October 21, OpenAI revealed that its "Patch the Planet" cybersecurity initiative had identified serious security weaknesses in PHP, the open-source language powering approximately 70% of web infrastructure. Working with GPT-5.6 Sol, researchers discovered the flaws and subsequently mobilized a coalition spanning PHP maintainers, security organizations, and government agencies to develop and deploy fixes across a two-month period. OpenAI attributed the extended timeline to the scope of the work, which demanded substantial modifications to the PHP engine itself.
The PHP Foundation has issued patched releases—versions 8.2.35, 8.3.36, 8.4.27, and 8.5.12—to address the vulnerabilities. Older versions, including PHP 8.1 and earlier, will not receive official patches. The coordinated disclosure was initially planned for October 22 but moved up to October 21 after unauthorized leaks on October 20 triggered public concern.
OpenAI characterizes this discovery as one of the most significant cybersecurity breakthroughs of 2026, given PHP's widespread deployment. The company frames the incident as a demonstration of its "Defender's Window" concept—a strategic window in which organizations can deploy advanced AI models to identify, remediate, and fortify defenses before threat actors gain access to comparable AI-powered attack tools. OpenAI developed this framework following an incident in which its models successfully breached Hugging Face.
Calif, a California-based security firm and founding partner of Patch the Planet, initiated the PHP audit using OpenAI tools on July 30, 2026, at the PHP Foundation's request. The Foundation had been aware of the underlying flaw since 2024 but lacked both an exploitation technique and a solution. Calif researcher Ryan A. pinpointed the exploitable vulnerability on August 11, with OpenAI researchers making the same discovery that same day. Wiz, another security organization, subsequently uncovered a second distinct vulnerability.
The Vulnerabilities
The two flaws, designated CVE-2026-109618 and CVE-2026-109619, stem from defects in PHP's memory management and error handling mechanisms. The first arises when PHP exhausts its allocated memory during request processing, potentially leaving incomplete internal structures that the engine may later incorrectly recognize as legitimate. The second involves a situation where a fatal error causes PHP to retain a pointer to memory that has been deallocated, enabling subsequent requests to read or modify that memory. Under specific circumstances, remote attackers—without requiring authentication—can leverage these weaknesses to run arbitrary code on vulnerable servers and potentially seize control of websites.
Through AI analysis, researchers also demonstrated that these flaws could leak sensitive data residing in server memory, such as passwords, session tokens, and API credentials. Particularly troubling is that both vulnerabilities reside in PHP's fundamental engine, meaning even websites with secure application code remain at risk. Each vulnerability received a maximum critical severity rating of 9.2 out of 10.
AI's Role in Discovery and Remediation
OpenAI notes that human developers and security professionals examined the affected code but were unable to fully resolve the issues, allowing them to persist for two years following initial detection. In contrast, researchers equipped with AI tools identified practical exploitation methods within fourteen days. Alternatively, one might contend that without AI involvement, these flaws would never have become exploitable in the first place, rendering them effectively benign.
OpenAI contests this perspective through its Defender's Window framework. The organization contends that transformative technology will inevitably be weaponized by malicious actors, and therefore society bears a shared obligation to proactively identify and eliminate any exploitable weaknesses. Anthropic's Claude Mythos has similarly reported discovering "thousands of zero-day vulnerabilities" across "every major operating system and every major web browser."